[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

GNUbg optimal cache size for analysis

Posted By: Philippe Michel
Date: Thursday, 19 May 2011, at 7:42 p.m.

In Response To: GNUbg optimal cache size for analysis (leobueno)

Wondering if this is a typical result, that is, for evaluations (as opposed to rollouts), *bigger* cache is indeed *better*.

I think it is, and it would apply to rollouts as well, with two reserves :

- it must be well within your machine's memory size (in your example, 336 MB is fine if you have 1+ GB of memory ; if you have 512 MB or less, it is too high)

- if what you do doesn't take much time, you will evaluate so few positions (relatively speaking) that the larger cache won't make a difference. Your example took 20-30 minutes. A 2-ply analysis taking 20-30 seconds may have gone : 32s 26s 25s 25s 25s 25s ... or something like that. This is a concern with the example given by Michael Petch in another answer.

with no neural net pruning

This, on the other hand, looks like a bad idea, or at least very expensive paranoia.

With pruning your analysis should be about twice as fast and the risk of discrepancies with no pruning is something like :

- less than 1% risk of 0.001 or higher difference (for any move or cube decision)

- less than 0.1% risk of 0.01 or higher difference

In terms of ER/PR, it should amount to 0.01-0.02. Rather marginal compared to the inaccuracy of the evaluation itself.

Messages In This Thread

 

Post Response
Your Name:
Your E-Mail Address:
Subject:
Message:

If necessary, enter your password below:

Password:
 

 

[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums is maintained by Stick with WebBBS 5.12.